2 * Block multiqueue core code
4 * Copyright (C) 2013-2014 Jens Axboe
5 * Copyright (C) 2013-2014 Christoph Hellwig
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/backing-dev.h>
10 #include <linux/bio.h>
11 #include <linux/blkdev.h>
13 #include <linux/init.h>
14 #include <linux/slab.h>
15 #include <linux/workqueue.h>
16 #include <linux/smp.h>
17 #include <linux/llist.h>
18 #include <linux/list_sort.h>
19 #include <linux/cpu.h>
20 #include <linux/cache.h>
21 #include <linux/sched/sysctl.h>
22 #include <linux/delay.h>
24 #include <trace/events/block.h>
26 #include <linux/blk-mq.h>
29 #include "blk-mq-tag.h"
31 static DEFINE_MUTEX(all_q_mutex);
32 static LIST_HEAD(all_q_list);
34 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
37 * Check if any of the ctx's have pending work in this hardware queue
39 static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
43 for (i = 0; i < hctx->ctx_map.map_size; i++)
44 if (hctx->ctx_map.map[i].word)
50 static inline struct blk_align_bitmap *get_bm(struct blk_mq_hw_ctx *hctx,
51 struct blk_mq_ctx *ctx)
53 return &hctx->ctx_map.map[ctx->index_hw / hctx->ctx_map.bits_per_word];
56 #define CTX_TO_BIT(hctx, ctx) \
57 ((ctx)->index_hw & ((hctx)->ctx_map.bits_per_word - 1))
60 * Mark this ctx as having pending work in this hardware queue
62 static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
63 struct blk_mq_ctx *ctx)
65 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
67 if (!test_bit(CTX_TO_BIT(hctx, ctx), &bm->word))
68 set_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
71 static void blk_mq_hctx_clear_pending(struct blk_mq_hw_ctx *hctx,
72 struct blk_mq_ctx *ctx)
74 struct blk_align_bitmap *bm = get_bm(hctx, ctx);
76 clear_bit(CTX_TO_BIT(hctx, ctx), &bm->word);
79 static int blk_mq_queue_enter(struct request_queue *q)
84 if (percpu_ref_tryget_live(&q->mq_usage_counter))
87 ret = wait_event_interruptible(q->mq_freeze_wq,
88 !q->mq_freeze_depth || blk_queue_dying(q));
89 if (blk_queue_dying(q))
96 static void blk_mq_queue_exit(struct request_queue *q)
98 percpu_ref_put(&q->mq_usage_counter);
101 static void blk_mq_usage_counter_release(struct percpu_ref *ref)
103 struct request_queue *q =
104 container_of(ref, struct request_queue, mq_usage_counter);
106 wake_up_all(&q->mq_freeze_wq);
110 * Guarantee no request is in use, so we can change any data structure of
111 * the queue afterward.
113 void blk_mq_freeze_queue(struct request_queue *q)
117 spin_lock_irq(q->queue_lock);
118 freeze = !q->mq_freeze_depth++;
119 spin_unlock_irq(q->queue_lock);
122 percpu_ref_kill(&q->mq_usage_counter);
123 blk_mq_run_queues(q, false);
125 wait_event(q->mq_freeze_wq, percpu_ref_is_zero(&q->mq_usage_counter));
128 static void blk_mq_unfreeze_queue(struct request_queue *q)
132 spin_lock_irq(q->queue_lock);
133 wake = !--q->mq_freeze_depth;
134 WARN_ON_ONCE(q->mq_freeze_depth < 0);
135 spin_unlock_irq(q->queue_lock);
137 percpu_ref_reinit(&q->mq_usage_counter);
138 wake_up_all(&q->mq_freeze_wq);
142 bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
144 return blk_mq_has_free_tags(hctx->tags);
146 EXPORT_SYMBOL(blk_mq_can_queue);
148 static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
149 struct request *rq, unsigned int rw_flags)
151 if (blk_queue_io_stat(q))
152 rw_flags |= REQ_IO_STAT;
154 INIT_LIST_HEAD(&rq->queuelist);
155 /* csd/requeue_work/fifo_time is initialized before use */
158 rq->cmd_flags |= rw_flags;
159 /* do not touch atomic flags, it needs atomic ops against the timer */
161 INIT_HLIST_NODE(&rq->hash);
162 RB_CLEAR_NODE(&rq->rb_node);
165 rq->start_time = jiffies;
166 #ifdef CONFIG_BLK_CGROUP
168 set_start_time_ns(rq);
169 rq->io_start_time_ns = 0;
171 rq->nr_phys_segments = 0;
172 #if defined(CONFIG_BLK_DEV_INTEGRITY)
173 rq->nr_integrity_segments = 0;
176 /* tag was already set */
186 INIT_LIST_HEAD(&rq->timeout_list);
190 rq->end_io_data = NULL;
193 ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
196 static struct request *
197 __blk_mq_alloc_request(struct blk_mq_alloc_data *data, int rw)
202 tag = blk_mq_get_tag(data);
203 if (tag != BLK_MQ_TAG_FAIL) {
204 rq = data->hctx->tags->rqs[tag];
206 if (blk_mq_tag_busy(data->hctx)) {
207 rq->cmd_flags = REQ_MQ_INFLIGHT;
208 atomic_inc(&data->hctx->nr_active);
212 blk_mq_rq_ctx_init(data->q, data->ctx, rq, rw);
219 struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp,
222 struct blk_mq_ctx *ctx;
223 struct blk_mq_hw_ctx *hctx;
225 struct blk_mq_alloc_data alloc_data;
228 ret = blk_mq_queue_enter(q);
232 ctx = blk_mq_get_ctx(q);
233 hctx = q->mq_ops->map_queue(q, ctx->cpu);
234 blk_mq_set_alloc_data(&alloc_data, q, gfp & ~__GFP_WAIT,
235 reserved, ctx, hctx);
237 rq = __blk_mq_alloc_request(&alloc_data, rw);
238 if (!rq && (gfp & __GFP_WAIT)) {
239 __blk_mq_run_hw_queue(hctx);
242 ctx = blk_mq_get_ctx(q);
243 hctx = q->mq_ops->map_queue(q, ctx->cpu);
244 blk_mq_set_alloc_data(&alloc_data, q, gfp, reserved, ctx,
246 rq = __blk_mq_alloc_request(&alloc_data, rw);
247 ctx = alloc_data.ctx;
251 return ERR_PTR(-EWOULDBLOCK);
254 EXPORT_SYMBOL(blk_mq_alloc_request);
256 static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
257 struct blk_mq_ctx *ctx, struct request *rq)
259 const int tag = rq->tag;
260 struct request_queue *q = rq->q;
262 if (rq->cmd_flags & REQ_MQ_INFLIGHT)
263 atomic_dec(&hctx->nr_active);
266 clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
267 blk_mq_put_tag(hctx, tag, &ctx->last_tag);
268 blk_mq_queue_exit(q);
271 void blk_mq_free_request(struct request *rq)
273 struct blk_mq_ctx *ctx = rq->mq_ctx;
274 struct blk_mq_hw_ctx *hctx;
275 struct request_queue *q = rq->q;
277 ctx->rq_completed[rq_is_sync(rq)]++;
279 hctx = q->mq_ops->map_queue(q, ctx->cpu);
280 __blk_mq_free_request(hctx, ctx, rq);
284 * Clone all relevant state from a request that has been put on hold in
285 * the flush state machine into the preallocated flush request that hangs
286 * off the request queue.
288 * For a driver the flush request should be invisible, that's why we are
289 * impersonating the original request here.
291 void blk_mq_clone_flush_request(struct request *flush_rq,
292 struct request *orig_rq)
294 struct blk_mq_hw_ctx *hctx =
295 orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
297 flush_rq->mq_ctx = orig_rq->mq_ctx;
298 flush_rq->tag = orig_rq->tag;
299 memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
303 inline void __blk_mq_end_request(struct request *rq, int error)
305 blk_account_io_done(rq);
308 rq->end_io(rq, error);
310 if (unlikely(blk_bidi_rq(rq)))
311 blk_mq_free_request(rq->next_rq);
312 blk_mq_free_request(rq);
315 EXPORT_SYMBOL(__blk_mq_end_request);
317 void blk_mq_end_request(struct request *rq, int error)
319 if (blk_update_request(rq, error, blk_rq_bytes(rq)))
321 __blk_mq_end_request(rq, error);
323 EXPORT_SYMBOL(blk_mq_end_request);
325 static void __blk_mq_complete_request_remote(void *data)
327 struct request *rq = data;
329 rq->q->softirq_done_fn(rq);
332 static void blk_mq_ipi_complete_request(struct request *rq)
334 struct blk_mq_ctx *ctx = rq->mq_ctx;
338 if (!test_bit(QUEUE_FLAG_SAME_COMP, &rq->q->queue_flags)) {
339 rq->q->softirq_done_fn(rq);
344 if (!test_bit(QUEUE_FLAG_SAME_FORCE, &rq->q->queue_flags))
345 shared = cpus_share_cache(cpu, ctx->cpu);
347 if (cpu != ctx->cpu && !shared && cpu_online(ctx->cpu)) {
348 rq->csd.func = __blk_mq_complete_request_remote;
351 smp_call_function_single_async(ctx->cpu, &rq->csd);
353 rq->q->softirq_done_fn(rq);
358 void __blk_mq_complete_request(struct request *rq)
360 struct request_queue *q = rq->q;
362 if (!q->softirq_done_fn)
363 blk_mq_end_request(rq, rq->errors);
365 blk_mq_ipi_complete_request(rq);
369 * blk_mq_complete_request - end I/O on a request
370 * @rq: the request being processed
373 * Ends all I/O on a request. It does not handle partial completions.
374 * The actual completion happens out-of-order, through a IPI handler.
376 void blk_mq_complete_request(struct request *rq)
378 struct request_queue *q = rq->q;
380 if (unlikely(blk_should_fake_timeout(q)))
382 if (!blk_mark_rq_complete(rq))
383 __blk_mq_complete_request(rq);
385 EXPORT_SYMBOL(blk_mq_complete_request);
387 void blk_mq_start_request(struct request *rq)
389 struct request_queue *q = rq->q;
391 trace_block_rq_issue(q, rq);
393 rq->resid_len = blk_rq_bytes(rq);
394 if (unlikely(blk_bidi_rq(rq)))
395 rq->next_rq->resid_len = blk_rq_bytes(rq->next_rq);
400 * Ensure that ->deadline is visible before set the started
401 * flag and clear the completed flag.
403 smp_mb__before_atomic();
406 * Mark us as started and clear complete. Complete might have been
407 * set if requeue raced with timeout, which then marked it as
408 * complete. So be sure to clear complete again when we start
409 * the request, otherwise we'll ignore the completion event.
411 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
412 set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
413 if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
414 clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
416 if (q->dma_drain_size && blk_rq_bytes(rq)) {
418 * Make sure space for the drain appears. We know we can do
419 * this because max_hw_segments has been adjusted to be one
420 * fewer than the device can handle.
422 rq->nr_phys_segments++;
425 EXPORT_SYMBOL(blk_mq_start_request);
427 static void __blk_mq_requeue_request(struct request *rq)
429 struct request_queue *q = rq->q;
431 trace_block_rq_requeue(q, rq);
433 if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
434 if (q->dma_drain_size && blk_rq_bytes(rq))
435 rq->nr_phys_segments--;
439 void blk_mq_requeue_request(struct request *rq)
441 __blk_mq_requeue_request(rq);
442 blk_clear_rq_complete(rq);
444 BUG_ON(blk_queued_rq(rq));
445 blk_mq_add_to_requeue_list(rq, true);
447 EXPORT_SYMBOL(blk_mq_requeue_request);
449 static void blk_mq_requeue_work(struct work_struct *work)
451 struct request_queue *q =
452 container_of(work, struct request_queue, requeue_work);
454 struct request *rq, *next;
457 spin_lock_irqsave(&q->requeue_lock, flags);
458 list_splice_init(&q->requeue_list, &rq_list);
459 spin_unlock_irqrestore(&q->requeue_lock, flags);
461 list_for_each_entry_safe(rq, next, &rq_list, queuelist) {
462 if (!(rq->cmd_flags & REQ_SOFTBARRIER))
465 rq->cmd_flags &= ~REQ_SOFTBARRIER;
466 list_del_init(&rq->queuelist);
467 blk_mq_insert_request(rq, true, false, false);
470 while (!list_empty(&rq_list)) {
471 rq = list_entry(rq_list.next, struct request, queuelist);
472 list_del_init(&rq->queuelist);
473 blk_mq_insert_request(rq, false, false, false);
477 * Use the start variant of queue running here, so that running
478 * the requeue work will kick stopped queues.
480 blk_mq_start_hw_queues(q);
483 void blk_mq_add_to_requeue_list(struct request *rq, bool at_head)
485 struct request_queue *q = rq->q;
489 * We abuse this flag that is otherwise used by the I/O scheduler to
490 * request head insertation from the workqueue.
492 BUG_ON(rq->cmd_flags & REQ_SOFTBARRIER);
494 spin_lock_irqsave(&q->requeue_lock, flags);
496 rq->cmd_flags |= REQ_SOFTBARRIER;
497 list_add(&rq->queuelist, &q->requeue_list);
499 list_add_tail(&rq->queuelist, &q->requeue_list);
501 spin_unlock_irqrestore(&q->requeue_lock, flags);
503 EXPORT_SYMBOL(blk_mq_add_to_requeue_list);
505 void blk_mq_kick_requeue_list(struct request_queue *q)
507 kblockd_schedule_work(&q->requeue_work);
509 EXPORT_SYMBOL(blk_mq_kick_requeue_list);
511 static inline bool is_flush_request(struct request *rq, unsigned int tag)
513 return ((rq->cmd_flags & REQ_FLUSH_SEQ) &&
514 rq->q->flush_rq->tag == tag);
517 struct request *blk_mq_tag_to_rq(struct blk_mq_tags *tags, unsigned int tag)
519 struct request *rq = tags->rqs[tag];
521 if (!is_flush_request(rq, tag))
524 return rq->q->flush_rq;
526 EXPORT_SYMBOL(blk_mq_tag_to_rq);
528 struct blk_mq_timeout_data {
530 unsigned int next_set;
533 static void blk_mq_rq_timed_out(struct request *req)
535 struct blk_mq_ops *ops = req->q->mq_ops;
536 enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
539 * We know that complete is set at this point. If STARTED isn't set
540 * anymore, then the request isn't active and the "timeout" should
541 * just be ignored. This can happen due to the bitflag ordering.
542 * Timeout first checks if STARTED is set, and if it is, assumes
543 * the request is active. But if we race with completion, then
544 * we both flags will get cleared. So check here again, and ignore
545 * a timeout event with a request that isn't active.
547 if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
551 ret = ops->timeout(req);
555 __blk_mq_complete_request(req);
557 case BLK_EH_RESET_TIMER:
559 blk_clear_rq_complete(req);
561 case BLK_EH_NOT_HANDLED:
564 printk(KERN_ERR "block: bad eh return: %d\n", ret);
569 static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
570 struct request *rq, void *priv, bool reserved)
572 struct blk_mq_timeout_data *data = priv;
574 if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
577 if (time_after_eq(jiffies, rq->deadline)) {
578 if (!blk_mark_rq_complete(rq))
579 blk_mq_rq_timed_out(rq);
580 } else if (!data->next_set || time_after(data->next, rq->deadline)) {
581 data->next = rq->deadline;
586 static void blk_mq_rq_timer(unsigned long priv)
588 struct request_queue *q = (struct request_queue *)priv;
589 struct blk_mq_timeout_data data = {
593 struct blk_mq_hw_ctx *hctx;
596 queue_for_each_hw_ctx(q, hctx, i) {
598 * If not software queues are currently mapped to this
599 * hardware queue, there's nothing to check
601 if (!hctx->nr_ctx || !hctx->tags)
604 blk_mq_tag_busy_iter(hctx, blk_mq_check_expired, &data);
608 data.next = blk_rq_timeout(round_jiffies_up(data.next));
609 mod_timer(&q->timeout, data.next);
611 queue_for_each_hw_ctx(q, hctx, i)
612 blk_mq_tag_idle(hctx);
617 * Reverse check our software queue for entries that we could potentially
618 * merge with. Currently includes a hand-wavy stop count of 8, to not spend
619 * too much time checking for merges.
621 static bool blk_mq_attempt_merge(struct request_queue *q,
622 struct blk_mq_ctx *ctx, struct bio *bio)
627 list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
633 if (!blk_rq_merge_ok(rq, bio))
636 el_ret = blk_try_merge(rq, bio);
637 if (el_ret == ELEVATOR_BACK_MERGE) {
638 if (bio_attempt_back_merge(q, rq, bio)) {
643 } else if (el_ret == ELEVATOR_FRONT_MERGE) {
644 if (bio_attempt_front_merge(q, rq, bio)) {
656 * Process software queues that have been marked busy, splicing them
657 * to the for-dispatch
659 static void flush_busy_ctxs(struct blk_mq_hw_ctx *hctx, struct list_head *list)
661 struct blk_mq_ctx *ctx;
664 for (i = 0; i < hctx->ctx_map.map_size; i++) {
665 struct blk_align_bitmap *bm = &hctx->ctx_map.map[i];
666 unsigned int off, bit;
672 off = i * hctx->ctx_map.bits_per_word;
674 bit = find_next_bit(&bm->word, bm->depth, bit);
675 if (bit >= bm->depth)
678 ctx = hctx->ctxs[bit + off];
679 clear_bit(bit, &bm->word);
680 spin_lock(&ctx->lock);
681 list_splice_tail_init(&ctx->rq_list, list);
682 spin_unlock(&ctx->lock);
690 * Run this hardware queue, pulling any software queues mapped to it in.
691 * Note that this function currently has various problems around ordering
692 * of IO. In particular, we'd like FIFO behaviour on handling existing
693 * items on the hctx->dispatch list. Ignore that for now.
695 static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
697 struct request_queue *q = hctx->queue;
702 WARN_ON(!cpumask_test_cpu(raw_smp_processor_id(), hctx->cpumask));
704 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
710 * Touch any software queue that has pending entries.
712 flush_busy_ctxs(hctx, &rq_list);
715 * If we have previous entries on our dispatch list, grab them
716 * and stuff them at the front for more fair dispatch.
718 if (!list_empty_careful(&hctx->dispatch)) {
719 spin_lock(&hctx->lock);
720 if (!list_empty(&hctx->dispatch))
721 list_splice_init(&hctx->dispatch, &rq_list);
722 spin_unlock(&hctx->lock);
726 * Now process all the entries, sending them to the driver.
729 while (!list_empty(&rq_list)) {
732 rq = list_first_entry(&rq_list, struct request, queuelist);
733 list_del_init(&rq->queuelist);
735 ret = q->mq_ops->queue_rq(hctx, rq, list_empty(&rq_list));
737 case BLK_MQ_RQ_QUEUE_OK:
740 case BLK_MQ_RQ_QUEUE_BUSY:
741 list_add(&rq->queuelist, &rq_list);
742 __blk_mq_requeue_request(rq);
745 pr_err("blk-mq: bad return on queue: %d\n", ret);
746 case BLK_MQ_RQ_QUEUE_ERROR:
748 blk_mq_end_request(rq, rq->errors);
752 if (ret == BLK_MQ_RQ_QUEUE_BUSY)
757 hctx->dispatched[0]++;
758 else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
759 hctx->dispatched[ilog2(queued) + 1]++;
762 * Any items that need requeuing? Stuff them into hctx->dispatch,
763 * that is where we will continue on next queue run.
765 if (!list_empty(&rq_list)) {
766 spin_lock(&hctx->lock);
767 list_splice(&rq_list, &hctx->dispatch);
768 spin_unlock(&hctx->lock);
773 * It'd be great if the workqueue API had a way to pass
774 * in a mask and had some smarts for more clever placement.
775 * For now we just round-robin here, switching for every
776 * BLK_MQ_CPU_WORK_BATCH queued items.
778 static int blk_mq_hctx_next_cpu(struct blk_mq_hw_ctx *hctx)
780 int cpu = hctx->next_cpu;
782 if (--hctx->next_cpu_batch <= 0) {
785 next_cpu = cpumask_next(hctx->next_cpu, hctx->cpumask);
786 if (next_cpu >= nr_cpu_ids)
787 next_cpu = cpumask_first(hctx->cpumask);
789 hctx->next_cpu = next_cpu;
790 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
796 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
798 if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
801 if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask))
802 __blk_mq_run_hw_queue(hctx);
803 else if (hctx->queue->nr_hw_queues == 1)
804 kblockd_schedule_delayed_work(&hctx->run_work, 0);
808 cpu = blk_mq_hctx_next_cpu(hctx);
809 kblockd_schedule_delayed_work_on(cpu, &hctx->run_work, 0);
813 void blk_mq_run_queues(struct request_queue *q, bool async)
815 struct blk_mq_hw_ctx *hctx;
818 queue_for_each_hw_ctx(q, hctx, i) {
819 if ((!blk_mq_hctx_has_pending(hctx) &&
820 list_empty_careful(&hctx->dispatch)) ||
821 test_bit(BLK_MQ_S_STOPPED, &hctx->state))
825 blk_mq_run_hw_queue(hctx, async);
829 EXPORT_SYMBOL(blk_mq_run_queues);
831 void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
833 cancel_delayed_work(&hctx->run_work);
834 cancel_delayed_work(&hctx->delay_work);
835 set_bit(BLK_MQ_S_STOPPED, &hctx->state);
837 EXPORT_SYMBOL(blk_mq_stop_hw_queue);
839 void blk_mq_stop_hw_queues(struct request_queue *q)
841 struct blk_mq_hw_ctx *hctx;
844 queue_for_each_hw_ctx(q, hctx, i)
845 blk_mq_stop_hw_queue(hctx);
847 EXPORT_SYMBOL(blk_mq_stop_hw_queues);
849 void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
851 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
854 blk_mq_run_hw_queue(hctx, false);
857 EXPORT_SYMBOL(blk_mq_start_hw_queue);
859 void blk_mq_start_hw_queues(struct request_queue *q)
861 struct blk_mq_hw_ctx *hctx;
864 queue_for_each_hw_ctx(q, hctx, i)
865 blk_mq_start_hw_queue(hctx);
867 EXPORT_SYMBOL(blk_mq_start_hw_queues);
870 void blk_mq_start_stopped_hw_queues(struct request_queue *q, bool async)
872 struct blk_mq_hw_ctx *hctx;
875 queue_for_each_hw_ctx(q, hctx, i) {
876 if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
879 clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
881 blk_mq_run_hw_queue(hctx, async);
885 EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
887 static void blk_mq_run_work_fn(struct work_struct *work)
889 struct blk_mq_hw_ctx *hctx;
891 hctx = container_of(work, struct blk_mq_hw_ctx, run_work.work);
893 __blk_mq_run_hw_queue(hctx);
896 static void blk_mq_delay_work_fn(struct work_struct *work)
898 struct blk_mq_hw_ctx *hctx;
900 hctx = container_of(work, struct blk_mq_hw_ctx, delay_work.work);
902 if (test_and_clear_bit(BLK_MQ_S_STOPPED, &hctx->state))
903 __blk_mq_run_hw_queue(hctx);
906 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
908 unsigned long tmo = msecs_to_jiffies(msecs);
910 if (hctx->queue->nr_hw_queues == 1)
911 kblockd_schedule_delayed_work(&hctx->delay_work, tmo);
915 cpu = blk_mq_hctx_next_cpu(hctx);
916 kblockd_schedule_delayed_work_on(cpu, &hctx->delay_work, tmo);
919 EXPORT_SYMBOL(blk_mq_delay_queue);
921 static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
922 struct request *rq, bool at_head)
924 struct blk_mq_ctx *ctx = rq->mq_ctx;
926 trace_block_rq_insert(hctx->queue, rq);
929 list_add(&rq->queuelist, &ctx->rq_list);
931 list_add_tail(&rq->queuelist, &ctx->rq_list);
933 blk_mq_hctx_mark_pending(hctx, ctx);
936 void blk_mq_insert_request(struct request *rq, bool at_head, bool run_queue,
939 struct request_queue *q = rq->q;
940 struct blk_mq_hw_ctx *hctx;
941 struct blk_mq_ctx *ctx = rq->mq_ctx, *current_ctx;
943 current_ctx = blk_mq_get_ctx(q);
944 if (!cpu_online(ctx->cpu))
945 rq->mq_ctx = ctx = current_ctx;
947 hctx = q->mq_ops->map_queue(q, ctx->cpu);
949 spin_lock(&ctx->lock);
950 __blk_mq_insert_request(hctx, rq, at_head);
951 spin_unlock(&ctx->lock);
954 blk_mq_run_hw_queue(hctx, async);
956 blk_mq_put_ctx(current_ctx);
959 static void blk_mq_insert_requests(struct request_queue *q,
960 struct blk_mq_ctx *ctx,
961 struct list_head *list,
966 struct blk_mq_hw_ctx *hctx;
967 struct blk_mq_ctx *current_ctx;
969 trace_block_unplug(q, depth, !from_schedule);
971 current_ctx = blk_mq_get_ctx(q);
973 if (!cpu_online(ctx->cpu))
975 hctx = q->mq_ops->map_queue(q, ctx->cpu);
978 * preemption doesn't flush plug list, so it's possible ctx->cpu is
981 spin_lock(&ctx->lock);
982 while (!list_empty(list)) {
985 rq = list_first_entry(list, struct request, queuelist);
986 list_del_init(&rq->queuelist);
988 __blk_mq_insert_request(hctx, rq, false);
990 spin_unlock(&ctx->lock);
992 blk_mq_run_hw_queue(hctx, from_schedule);
993 blk_mq_put_ctx(current_ctx);
996 static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
998 struct request *rqa = container_of(a, struct request, queuelist);
999 struct request *rqb = container_of(b, struct request, queuelist);
1001 return !(rqa->mq_ctx < rqb->mq_ctx ||
1002 (rqa->mq_ctx == rqb->mq_ctx &&
1003 blk_rq_pos(rqa) < blk_rq_pos(rqb)));
1006 void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
1008 struct blk_mq_ctx *this_ctx;
1009 struct request_queue *this_q;
1012 LIST_HEAD(ctx_list);
1015 list_splice_init(&plug->mq_list, &list);
1017 list_sort(NULL, &list, plug_ctx_cmp);
1023 while (!list_empty(&list)) {
1024 rq = list_entry_rq(list.next);
1025 list_del_init(&rq->queuelist);
1027 if (rq->mq_ctx != this_ctx) {
1029 blk_mq_insert_requests(this_q, this_ctx,
1034 this_ctx = rq->mq_ctx;
1040 list_add_tail(&rq->queuelist, &ctx_list);
1044 * If 'this_ctx' is set, we know we have entries to complete
1045 * on 'ctx_list'. Do those.
1048 blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
1053 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
1055 init_request_from_bio(rq, bio);
1057 if (blk_do_io_stat(rq))
1058 blk_account_io_start(rq, 1);
1061 static inline bool hctx_allow_merges(struct blk_mq_hw_ctx *hctx)
1063 return (hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
1064 !blk_queue_nomerges(hctx->queue);
1067 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,
1068 struct blk_mq_ctx *ctx,
1069 struct request *rq, struct bio *bio)
1071 if (!hctx_allow_merges(hctx)) {
1072 blk_mq_bio_to_request(rq, bio);
1073 spin_lock(&ctx->lock);
1075 __blk_mq_insert_request(hctx, rq, false);
1076 spin_unlock(&ctx->lock);
1079 struct request_queue *q = hctx->queue;
1081 spin_lock(&ctx->lock);
1082 if (!blk_mq_attempt_merge(q, ctx, bio)) {
1083 blk_mq_bio_to_request(rq, bio);
1087 spin_unlock(&ctx->lock);
1088 __blk_mq_free_request(hctx, ctx, rq);
1093 struct blk_map_ctx {
1094 struct blk_mq_hw_ctx *hctx;
1095 struct blk_mq_ctx *ctx;
1098 static struct request *blk_mq_map_request(struct request_queue *q,
1100 struct blk_map_ctx *data)
1102 struct blk_mq_hw_ctx *hctx;
1103 struct blk_mq_ctx *ctx;
1105 int rw = bio_data_dir(bio);
1106 struct blk_mq_alloc_data alloc_data;
1108 if (unlikely(blk_mq_queue_enter(q))) {
1109 bio_endio(bio, -EIO);
1113 ctx = blk_mq_get_ctx(q);
1114 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1116 if (rw_is_sync(bio->bi_rw))
1119 trace_block_getrq(q, bio, rw);
1120 blk_mq_set_alloc_data(&alloc_data, q, GFP_ATOMIC, false, ctx,
1122 rq = __blk_mq_alloc_request(&alloc_data, rw);
1123 if (unlikely(!rq)) {
1124 __blk_mq_run_hw_queue(hctx);
1125 blk_mq_put_ctx(ctx);
1126 trace_block_sleeprq(q, bio, rw);
1128 ctx = blk_mq_get_ctx(q);
1129 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1130 blk_mq_set_alloc_data(&alloc_data, q,
1131 __GFP_WAIT|GFP_ATOMIC, false, ctx, hctx);
1132 rq = __blk_mq_alloc_request(&alloc_data, rw);
1133 ctx = alloc_data.ctx;
1134 hctx = alloc_data.hctx;
1144 * Multiple hardware queue variant. This will not use per-process plugs,
1145 * but will attempt to bypass the hctx queueing if we can go straight to
1146 * hardware for SYNC IO.
1148 static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
1150 const int is_sync = rw_is_sync(bio->bi_rw);
1151 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1152 struct blk_map_ctx data;
1155 blk_queue_bounce(q, &bio);
1157 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1158 bio_endio(bio, -EIO);
1162 rq = blk_mq_map_request(q, bio, &data);
1166 if (unlikely(is_flush_fua)) {
1167 blk_mq_bio_to_request(rq, bio);
1168 blk_insert_flush(rq);
1175 blk_mq_bio_to_request(rq, bio);
1178 * For OK queue, we are done. For error, kill it. Any other
1179 * error (busy), just add it to our list as we previously
1182 ret = q->mq_ops->queue_rq(data.hctx, rq, true);
1183 if (ret == BLK_MQ_RQ_QUEUE_OK)
1186 __blk_mq_requeue_request(rq);
1188 if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
1190 blk_mq_end_request(rq, rq->errors);
1196 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1198 * For a SYNC request, send it to the hardware immediately. For
1199 * an ASYNC request, just ensure that we run it later on. The
1200 * latter allows for merging opportunities and more efficient
1204 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1207 blk_mq_put_ctx(data.ctx);
1211 * Single hardware queue variant. This will attempt to use any per-process
1212 * plug for merging and IO deferral.
1214 static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
1216 const int is_sync = rw_is_sync(bio->bi_rw);
1217 const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
1218 unsigned int use_plug, request_count = 0;
1219 struct blk_map_ctx data;
1223 * If we have multiple hardware queues, just go directly to
1224 * one of those for sync IO.
1226 use_plug = !is_flush_fua && !is_sync;
1228 blk_queue_bounce(q, &bio);
1230 if (bio_integrity_enabled(bio) && bio_integrity_prep(bio)) {
1231 bio_endio(bio, -EIO);
1235 if (use_plug && !blk_queue_nomerges(q) &&
1236 blk_attempt_plug_merge(q, bio, &request_count))
1239 rq = blk_mq_map_request(q, bio, &data);
1243 if (unlikely(is_flush_fua)) {
1244 blk_mq_bio_to_request(rq, bio);
1245 blk_insert_flush(rq);
1250 * A task plug currently exists. Since this is completely lockless,
1251 * utilize that to temporarily store requests until the task is
1252 * either done or scheduled away.
1255 struct blk_plug *plug = current->plug;
1258 blk_mq_bio_to_request(rq, bio);
1259 if (list_empty(&plug->mq_list))
1260 trace_block_plug(q);
1261 else if (request_count >= BLK_MAX_REQUEST_COUNT) {
1262 blk_flush_plug_list(plug, false);
1263 trace_block_plug(q);
1265 list_add_tail(&rq->queuelist, &plug->mq_list);
1266 blk_mq_put_ctx(data.ctx);
1271 if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
1273 * For a SYNC request, send it to the hardware immediately. For
1274 * an ASYNC request, just ensure that we run it later on. The
1275 * latter allows for merging opportunities and more efficient
1279 blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
1282 blk_mq_put_ctx(data.ctx);
1286 * Default mapping to a software queue, since we use one per CPU.
1288 struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
1290 return q->queue_hw_ctx[q->mq_map[cpu]];
1292 EXPORT_SYMBOL(blk_mq_map_queue);
1294 static void blk_mq_free_rq_map(struct blk_mq_tag_set *set,
1295 struct blk_mq_tags *tags, unsigned int hctx_idx)
1299 if (tags->rqs && set->ops->exit_request) {
1302 for (i = 0; i < tags->nr_tags; i++) {
1305 set->ops->exit_request(set->driver_data, tags->rqs[i],
1307 tags->rqs[i] = NULL;
1311 while (!list_empty(&tags->page_list)) {
1312 page = list_first_entry(&tags->page_list, struct page, lru);
1313 list_del_init(&page->lru);
1314 __free_pages(page, page->private);
1319 blk_mq_free_tags(tags);
1322 static size_t order_to_size(unsigned int order)
1324 return (size_t)PAGE_SIZE << order;
1327 static struct blk_mq_tags *blk_mq_init_rq_map(struct blk_mq_tag_set *set,
1328 unsigned int hctx_idx)
1330 struct blk_mq_tags *tags;
1331 unsigned int i, j, entries_per_page, max_order = 4;
1332 size_t rq_size, left;
1334 tags = blk_mq_init_tags(set->queue_depth, set->reserved_tags,
1339 INIT_LIST_HEAD(&tags->page_list);
1341 tags->rqs = kzalloc_node(set->queue_depth * sizeof(struct request *),
1342 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1345 blk_mq_free_tags(tags);
1350 * rq_size is the size of the request plus driver payload, rounded
1351 * to the cacheline size
1353 rq_size = round_up(sizeof(struct request) + set->cmd_size,
1355 left = rq_size * set->queue_depth;
1357 for (i = 0; i < set->queue_depth; ) {
1358 int this_order = max_order;
1363 while (left < order_to_size(this_order - 1) && this_order)
1367 page = alloc_pages_node(set->numa_node,
1368 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY,
1374 if (order_to_size(this_order) < rq_size)
1381 page->private = this_order;
1382 list_add_tail(&page->lru, &tags->page_list);
1384 p = page_address(page);
1385 entries_per_page = order_to_size(this_order) / rq_size;
1386 to_do = min(entries_per_page, set->queue_depth - i);
1387 left -= to_do * rq_size;
1388 for (j = 0; j < to_do; j++) {
1390 tags->rqs[i]->atomic_flags = 0;
1391 tags->rqs[i]->cmd_flags = 0;
1392 if (set->ops->init_request) {
1393 if (set->ops->init_request(set->driver_data,
1394 tags->rqs[i], hctx_idx, i,
1396 tags->rqs[i] = NULL;
1409 blk_mq_free_rq_map(set, tags, hctx_idx);
1413 static void blk_mq_free_bitmap(struct blk_mq_ctxmap *bitmap)
1418 static int blk_mq_alloc_bitmap(struct blk_mq_ctxmap *bitmap, int node)
1420 unsigned int bpw = 8, total, num_maps, i;
1422 bitmap->bits_per_word = bpw;
1424 num_maps = ALIGN(nr_cpu_ids, bpw) / bpw;
1425 bitmap->map = kzalloc_node(num_maps * sizeof(struct blk_align_bitmap),
1430 bitmap->map_size = num_maps;
1433 for (i = 0; i < num_maps; i++) {
1434 bitmap->map[i].depth = min(total, bitmap->bits_per_word);
1435 total -= bitmap->map[i].depth;
1441 static int blk_mq_hctx_cpu_offline(struct blk_mq_hw_ctx *hctx, int cpu)
1443 struct request_queue *q = hctx->queue;
1444 struct blk_mq_ctx *ctx;
1448 * Move ctx entries to new CPU, if this one is going away.
1450 ctx = __blk_mq_get_ctx(q, cpu);
1452 spin_lock(&ctx->lock);
1453 if (!list_empty(&ctx->rq_list)) {
1454 list_splice_init(&ctx->rq_list, &tmp);
1455 blk_mq_hctx_clear_pending(hctx, ctx);
1457 spin_unlock(&ctx->lock);
1459 if (list_empty(&tmp))
1462 ctx = blk_mq_get_ctx(q);
1463 spin_lock(&ctx->lock);
1465 while (!list_empty(&tmp)) {
1468 rq = list_first_entry(&tmp, struct request, queuelist);
1470 list_move_tail(&rq->queuelist, &ctx->rq_list);
1473 hctx = q->mq_ops->map_queue(q, ctx->cpu);
1474 blk_mq_hctx_mark_pending(hctx, ctx);
1476 spin_unlock(&ctx->lock);
1478 blk_mq_run_hw_queue(hctx, true);
1479 blk_mq_put_ctx(ctx);
1483 static int blk_mq_hctx_cpu_online(struct blk_mq_hw_ctx *hctx, int cpu)
1485 struct request_queue *q = hctx->queue;
1486 struct blk_mq_tag_set *set = q->tag_set;
1488 if (set->tags[hctx->queue_num])
1491 set->tags[hctx->queue_num] = blk_mq_init_rq_map(set, hctx->queue_num);
1492 if (!set->tags[hctx->queue_num])
1495 hctx->tags = set->tags[hctx->queue_num];
1499 static int blk_mq_hctx_notify(void *data, unsigned long action,
1502 struct blk_mq_hw_ctx *hctx = data;
1504 if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
1505 return blk_mq_hctx_cpu_offline(hctx, cpu);
1506 else if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN)
1507 return blk_mq_hctx_cpu_online(hctx, cpu);
1512 static void blk_mq_exit_hw_queues(struct request_queue *q,
1513 struct blk_mq_tag_set *set, int nr_queue)
1515 struct blk_mq_hw_ctx *hctx;
1518 queue_for_each_hw_ctx(q, hctx, i) {
1522 blk_mq_tag_idle(hctx);
1524 if (set->ops->exit_hctx)
1525 set->ops->exit_hctx(hctx, i);
1527 blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
1529 blk_mq_free_bitmap(&hctx->ctx_map);
1534 static void blk_mq_free_hw_queues(struct request_queue *q,
1535 struct blk_mq_tag_set *set)
1537 struct blk_mq_hw_ctx *hctx;
1540 queue_for_each_hw_ctx(q, hctx, i) {
1541 free_cpumask_var(hctx->cpumask);
1546 static int blk_mq_init_hw_queues(struct request_queue *q,
1547 struct blk_mq_tag_set *set)
1549 struct blk_mq_hw_ctx *hctx;
1553 * Initialize hardware queues
1555 queue_for_each_hw_ctx(q, hctx, i) {
1558 node = hctx->numa_node;
1559 if (node == NUMA_NO_NODE)
1560 node = hctx->numa_node = set->numa_node;
1562 INIT_DELAYED_WORK(&hctx->run_work, blk_mq_run_work_fn);
1563 INIT_DELAYED_WORK(&hctx->delay_work, blk_mq_delay_work_fn);
1564 spin_lock_init(&hctx->lock);
1565 INIT_LIST_HEAD(&hctx->dispatch);
1567 hctx->queue_num = i;
1568 hctx->flags = set->flags;
1569 hctx->cmd_size = set->cmd_size;
1571 blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
1572 blk_mq_hctx_notify, hctx);
1573 blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
1575 hctx->tags = set->tags[i];
1578 * Allocate space for all possible cpus to avoid allocation at
1581 hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
1586 if (blk_mq_alloc_bitmap(&hctx->ctx_map, node))
1591 if (set->ops->init_hctx &&
1592 set->ops->init_hctx(hctx, set->driver_data, i))
1596 if (i == q->nr_hw_queues)
1602 blk_mq_exit_hw_queues(q, set, i);
1607 static void blk_mq_init_cpu_queues(struct request_queue *q,
1608 unsigned int nr_hw_queues)
1612 for_each_possible_cpu(i) {
1613 struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
1614 struct blk_mq_hw_ctx *hctx;
1616 memset(__ctx, 0, sizeof(*__ctx));
1618 spin_lock_init(&__ctx->lock);
1619 INIT_LIST_HEAD(&__ctx->rq_list);
1622 /* If the cpu isn't online, the cpu is mapped to first hctx */
1626 hctx = q->mq_ops->map_queue(q, i);
1627 cpumask_set_cpu(i, hctx->cpumask);
1631 * Set local node, IFF we have more than one hw queue. If
1632 * not, we remain on the home node of the device
1634 if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
1635 hctx->numa_node = cpu_to_node(i);
1639 static void blk_mq_map_swqueue(struct request_queue *q)
1642 struct blk_mq_hw_ctx *hctx;
1643 struct blk_mq_ctx *ctx;
1645 queue_for_each_hw_ctx(q, hctx, i) {
1646 cpumask_clear(hctx->cpumask);
1651 * Map software to hardware queues
1653 queue_for_each_ctx(q, ctx, i) {
1654 /* If the cpu isn't online, the cpu is mapped to first hctx */
1658 hctx = q->mq_ops->map_queue(q, i);
1659 cpumask_set_cpu(i, hctx->cpumask);
1660 ctx->index_hw = hctx->nr_ctx;
1661 hctx->ctxs[hctx->nr_ctx++] = ctx;
1664 queue_for_each_hw_ctx(q, hctx, i) {
1666 * If no software queues are mapped to this hardware queue,
1667 * disable it and free the request entries.
1669 if (!hctx->nr_ctx) {
1670 struct blk_mq_tag_set *set = q->tag_set;
1673 blk_mq_free_rq_map(set, set->tags[i], i);
1674 set->tags[i] = NULL;
1681 * Initialize batch roundrobin counts
1683 hctx->next_cpu = cpumask_first(hctx->cpumask);
1684 hctx->next_cpu_batch = BLK_MQ_CPU_WORK_BATCH;
1688 static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set)
1690 struct blk_mq_hw_ctx *hctx;
1691 struct request_queue *q;
1695 if (set->tag_list.next == set->tag_list.prev)
1700 list_for_each_entry(q, &set->tag_list, tag_set_list) {
1701 blk_mq_freeze_queue(q);
1703 queue_for_each_hw_ctx(q, hctx, i) {
1705 hctx->flags |= BLK_MQ_F_TAG_SHARED;
1707 hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
1709 blk_mq_unfreeze_queue(q);
1713 static void blk_mq_del_queue_tag_set(struct request_queue *q)
1715 struct blk_mq_tag_set *set = q->tag_set;
1717 mutex_lock(&set->tag_list_lock);
1718 list_del_init(&q->tag_set_list);
1719 blk_mq_update_tag_set_depth(set);
1720 mutex_unlock(&set->tag_list_lock);
1723 static void blk_mq_add_queue_tag_set(struct blk_mq_tag_set *set,
1724 struct request_queue *q)
1728 mutex_lock(&set->tag_list_lock);
1729 list_add_tail(&q->tag_set_list, &set->tag_list);
1730 blk_mq_update_tag_set_depth(set);
1731 mutex_unlock(&set->tag_list_lock);
1734 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
1736 struct blk_mq_hw_ctx **hctxs;
1737 struct blk_mq_ctx __percpu *ctx;
1738 struct request_queue *q;
1742 ctx = alloc_percpu(struct blk_mq_ctx);
1744 return ERR_PTR(-ENOMEM);
1746 hctxs = kmalloc_node(set->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
1752 map = blk_mq_make_queue_map(set);
1756 for (i = 0; i < set->nr_hw_queues; i++) {
1757 int node = blk_mq_hw_queue_to_node(map, i);
1759 hctxs[i] = kzalloc_node(sizeof(struct blk_mq_hw_ctx),
1764 if (!zalloc_cpumask_var(&hctxs[i]->cpumask, GFP_KERNEL))
1767 atomic_set(&hctxs[i]->nr_active, 0);
1768 hctxs[i]->numa_node = node;
1769 hctxs[i]->queue_num = i;
1772 q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
1776 if (percpu_ref_init(&q->mq_usage_counter, blk_mq_usage_counter_release))
1779 setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
1780 blk_queue_rq_timeout(q, 30000);
1782 q->nr_queues = nr_cpu_ids;
1783 q->nr_hw_queues = set->nr_hw_queues;
1787 q->queue_hw_ctx = hctxs;
1789 q->mq_ops = set->ops;
1790 q->queue_flags |= QUEUE_FLAG_MQ_DEFAULT;
1792 if (!(set->flags & BLK_MQ_F_SG_MERGE))
1793 q->queue_flags |= 1 << QUEUE_FLAG_NO_SG_MERGE;
1795 q->sg_reserved_size = INT_MAX;
1797 INIT_WORK(&q->requeue_work, blk_mq_requeue_work);
1798 INIT_LIST_HEAD(&q->requeue_list);
1799 spin_lock_init(&q->requeue_lock);
1801 if (q->nr_hw_queues > 1)
1802 blk_queue_make_request(q, blk_mq_make_request);
1804 blk_queue_make_request(q, blk_sq_make_request);
1807 blk_queue_rq_timeout(q, set->timeout);
1810 * Do this after blk_queue_make_request() overrides it...
1812 q->nr_requests = set->queue_depth;
1814 if (set->ops->complete)
1815 blk_queue_softirq_done(q, set->ops->complete);
1817 blk_mq_init_flush(q);
1818 blk_mq_init_cpu_queues(q, set->nr_hw_queues);
1820 q->flush_rq = kzalloc(round_up(sizeof(struct request) +
1821 set->cmd_size, cache_line_size()),
1826 if (blk_mq_init_hw_queues(q, set))
1829 mutex_lock(&all_q_mutex);
1830 list_add_tail(&q->all_q_node, &all_q_list);
1831 mutex_unlock(&all_q_mutex);
1833 blk_mq_add_queue_tag_set(set, q);
1835 blk_mq_map_swqueue(q);
1842 blk_cleanup_queue(q);
1845 for (i = 0; i < set->nr_hw_queues; i++) {
1848 free_cpumask_var(hctxs[i]->cpumask);
1855 return ERR_PTR(-ENOMEM);
1857 EXPORT_SYMBOL(blk_mq_init_queue);
1859 void blk_mq_free_queue(struct request_queue *q)
1861 struct blk_mq_tag_set *set = q->tag_set;
1863 blk_mq_del_queue_tag_set(q);
1865 blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
1866 blk_mq_free_hw_queues(q, set);
1868 percpu_ref_exit(&q->mq_usage_counter);
1870 free_percpu(q->queue_ctx);
1871 kfree(q->queue_hw_ctx);
1874 q->queue_ctx = NULL;
1875 q->queue_hw_ctx = NULL;
1878 mutex_lock(&all_q_mutex);
1879 list_del_init(&q->all_q_node);
1880 mutex_unlock(&all_q_mutex);
1883 /* Basically redo blk_mq_init_queue with queue frozen */
1884 static void blk_mq_queue_reinit(struct request_queue *q)
1886 blk_mq_freeze_queue(q);
1888 blk_mq_sysfs_unregister(q);
1890 blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
1893 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
1894 * we should change hctx numa_node according to new topology (this
1895 * involves free and re-allocate memory, worthy doing?)
1898 blk_mq_map_swqueue(q);
1900 blk_mq_sysfs_register(q);
1902 blk_mq_unfreeze_queue(q);
1905 static int blk_mq_queue_reinit_notify(struct notifier_block *nb,
1906 unsigned long action, void *hcpu)
1908 struct request_queue *q;
1911 * Before new mappings are established, hotadded cpu might already
1912 * start handling requests. This doesn't break anything as we map
1913 * offline CPUs to first hardware queue. We will re-init the queue
1914 * below to get optimal settings.
1916 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
1917 action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
1920 mutex_lock(&all_q_mutex);
1921 list_for_each_entry(q, &all_q_list, all_q_node)
1922 blk_mq_queue_reinit(q);
1923 mutex_unlock(&all_q_mutex);
1927 static int __blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1931 for (i = 0; i < set->nr_hw_queues; i++) {
1932 set->tags[i] = blk_mq_init_rq_map(set, i);
1941 blk_mq_free_rq_map(set, set->tags[i], i);
1947 * Allocate the request maps associated with this tag_set. Note that this
1948 * may reduce the depth asked for, if memory is tight. set->queue_depth
1949 * will be updated to reflect the allocated depth.
1951 static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
1956 depth = set->queue_depth;
1958 err = __blk_mq_alloc_rq_maps(set);
1962 set->queue_depth >>= 1;
1963 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN) {
1967 } while (set->queue_depth);
1969 if (!set->queue_depth || err) {
1970 pr_err("blk-mq: failed to allocate request map\n");
1974 if (depth != set->queue_depth)
1975 pr_info("blk-mq: reduced tag depth (%u -> %u)\n",
1976 depth, set->queue_depth);
1982 * Alloc a tag set to be associated with one or more request queues.
1983 * May fail with EINVAL for various error conditions. May adjust the
1984 * requested depth down, if if it too large. In that case, the set
1985 * value will be stored in set->queue_depth.
1987 int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
1989 if (!set->nr_hw_queues)
1991 if (!set->queue_depth)
1993 if (set->queue_depth < set->reserved_tags + BLK_MQ_TAG_MIN)
1996 if (!set->nr_hw_queues || !set->ops->queue_rq || !set->ops->map_queue)
1999 if (set->queue_depth > BLK_MQ_MAX_DEPTH) {
2000 pr_info("blk-mq: reduced tag depth to %u\n",
2002 set->queue_depth = BLK_MQ_MAX_DEPTH;
2005 set->tags = kmalloc_node(set->nr_hw_queues *
2006 sizeof(struct blk_mq_tags *),
2007 GFP_KERNEL, set->numa_node);
2011 if (blk_mq_alloc_rq_maps(set))
2014 mutex_init(&set->tag_list_lock);
2015 INIT_LIST_HEAD(&set->tag_list);
2023 EXPORT_SYMBOL(blk_mq_alloc_tag_set);
2025 void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
2029 for (i = 0; i < set->nr_hw_queues; i++) {
2031 blk_mq_free_rq_map(set, set->tags[i], i);
2037 EXPORT_SYMBOL(blk_mq_free_tag_set);
2039 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
2041 struct blk_mq_tag_set *set = q->tag_set;
2042 struct blk_mq_hw_ctx *hctx;
2045 if (!set || nr > set->queue_depth)
2049 queue_for_each_hw_ctx(q, hctx, i) {
2050 ret = blk_mq_tag_update_depth(hctx->tags, nr);
2056 q->nr_requests = nr;
2061 void blk_mq_disable_hotplug(void)
2063 mutex_lock(&all_q_mutex);
2066 void blk_mq_enable_hotplug(void)
2068 mutex_unlock(&all_q_mutex);
2071 static int __init blk_mq_init(void)
2075 hotcpu_notifier(blk_mq_queue_reinit_notify, 0);
2079 subsys_initcall(blk_mq_init);